Professional Documents
Culture Documents
Patrik Fagerfjäll
Patrik Fagerfjäll
The two models are capable of optimizing the problems stated at a reasonable
computation time. The Production model was compared to the commercial
WindPRO 2.6 Optimize module in verification examples housing 20 to 30
wind turbines. For the problem of maximizing the total production in the
given geographic areas, the Production model of this thesis managed to find
locations for many more wind turbines than Optimize, yielding about 40%
higher total production. When restricted to allow only as many wind tur-
bines as the software Optimize was able to place, the Production model still
performed equal or better.
Our tests show that the field of mixed integer linear programming and the
mathematical models of this thesis possess a great potential to aid in the
process of increasing the return of wind farms. These tools are versatile and
can be adopted to many different scenarios. They need further development
to be able to handle really large wind farm projects but are still, as of today,
capable of delivering more bang for the buck in wind farm layout design.
i
Sammanfattning
Layouten hos en vindkraftpark designas idag ofta för hand. Som underlag
har man ett projektområde och utifrån en karta över terräng eller beräknad
energitäthet skall man placera ut vindturbiner för att nå en hög total pro-
duktion. En liten ändring av en vindturbins medelvind ger stora skillnader
i effektuttag, och en effektiv inbördes placering är därmed viktigt för att få
bra avkastning från parken. Lokala variationer i vindstyrka och vindfrekvens i
olika riktningar för olika delar av ett parkområde medför mycket information
att hålla reda på. Lägg till detta hur vakeffekterna mellan två vindturbiner
påverkar dessas produktion negativt samt hur ljudnivåer från flera vindtur-
biner samverkar till att öka minsta tillåtna avståndet till bebyggelse, och
processen blir väldigt detaljrik. Utöver dessa produktionsaspekter innefattar
designen också problem som att välja rätt dimension av kablar samt att dra
dessa och interna vägar på billigast möjliga sätt.
Två modeller skapas, av typen linjära heltalsprogram, där den ena repre-
senterar problemet att maximera produktion och den andra problemet att
minimera kostnader för kablar, vägnät och placering av transformatorsta-
tion. Båda modellerna kan beskrivas med hjälp av grafer, vilka består av en
mängd geografiskt utspridda noder samt bågar som sammanbinder dessa. No-
derna representerar möjliga placeringar av vindturbiner, där en nod i vilken
en vindturbin placerats kallas aktiv, och bågarna symboliserar vakeffekter,
kablar eller vägar mellan noder. En aktiv båge i Infrastruktursmodellen kan
exempelvis motsvara en kabeldragning mellan två vindturbiner.
ii
I Produktionsmodellen implementeras ett minsta tillåtna avstånd mellan två
vindturbiner som ett tätpackningsproblem av cirklar vars radier är halva det-
ta minsta avstånd. Två placerade vindturbiner medför produktionsförluster i
form av vakeffekter vilket representeras av en aktiv båge mellan motsvaran-
de noder i grafen. Den kombinerade effekten av vakeffekter från flera källor
beräknas genom linjär överlagring. Då andra kombinationsmetoder tidigare
har visat sig stämma bättre överens med verkligheten, exempelvis att sum-
mera kvadraterna av hastighetsförlusterna, skulle detta kunna vålla problem
för Produktionsmodellen, och en metodik för att i dessa modeller approxi-
mera andra samband än linjär överlagring presenteras. Vidare kompletteras
modellen med bivillkor som begränsar den kombinerade ljudnivån i en viss
punkt från alla vindturbiner och för olika vindriktningar. Detta innebär att
det minsta avståndet från en vindturbin till en närliggande bostad beror av
antalet vindturbiner, i vilka riktningar från bostaden de är placerade samt
hur vindkarakteristiken för dessa olika riktningar ser ut. Denna modellering
medför större möjligheter än om man endast anger ett minsta avstånd till
bebyggelse, vilket görs i Optimize.
iii
ten med vilken Produktionsmodellen kan lösa problem med hög upplösning.
Denna höga upplösning har mindre betydelse för storleken på vakeffekterna
än för möjligheten till precis och tät placering av vindturbinerna. För att
antalet bågvariabler i grafen därför inte skall växa som kvadraten på antalet
noder, föreslås en uppdelning av ytan i sektorer. Vindturbiner kan placeras
noggrant över det högupplösta rutnätet, medan vakeffekter samt eventuellt
kablar och vägar approximeras med bågar mellan de sektorer som innehåller
vindturbiner. Antalet bågar motsvarar då antalet sektorer i kvadrat obero-
ende av upplösningen. Storleken på dessa sektorer väljs så att det minsta
tillåtna avståndet mellan vindturbiner förbjuder fler än en vindturbin per
sektor samt så att vakeffekterna mellan sektorer förblir en god approxima-
tion av vakeffekterna mellan de riktiga positionerna för vindturbinerna.
De utvecklade modellerna visar sig fungera väl för sina respektive ändamål
och kan användas i processen att skapa en layout för en vindkraftpark. Pro-
duktionsmodellen kan optimera vindturbinernas placeringar med avseende på
antingen produktion eller vinst, och kan därmed bidra till en högre total pro-
duktion och större intäkter. Infrastrukturmodellen hittar prisvärda lösningar
på kabeldragningar med val av kabeltyp samt vägnät, vilket kan bidra till
att minska kostnaderna för dessa. Då båda modellerna kan hantera mycket
av den ingående problematiken kan också en större del av layoutprocessen
automatiseras och därmed tiden för det manuella arbetet minskas. Kombi-
nationen av de båda modellerna till en enda optimeringsmodell är intressant
ifall kostnaderna för kabel- och vägnät är av betydande storlek relativt vins-
ten från vindturbinernas produktion, då förbättringar hos dessa annars en-
dast leder till obetydliga förbättringar i totala vinsten och vindturbinernas
position sannolikt inte påverkas. Relationen mellan produktionsvinsten och
kostnaden för kabel- och vägnät beror på elpris samt vilken önskad återbetal-
ningstid för parken som används. Den kombinerade modellen blir dock stor
och tar lång tid att lösa och vissa föreslagna genvägar för att minska antalet
noder i infrastrukturmodellen inte kan användas.
iv
dock på olika sätt: Produktionsmodellen som överlagrar enskilda vakeffekter
överskattar den kombinerade förlusten från två vakar jämfört med Optimize
och PARK som båda kombinerar vakar som summan av kvadraterna av has-
tighetsförlusterna. Då Optimize inte kan hantera överlagring av ljudintensitet
från flera vindturbiner, utan endast använder ett minsta tillåtna avstånd från
var och en av dessa, skalades Produktionsmodellen ner till att inte inklude-
ra denna del. Jämförelsen visade att Produktionsmodellen finner en högre
total produktion än Optimize. Detta då den kunde placera ut många fler
vindturbiner än vad Optimize lyckades med. I fallet med en upplösning om
100m lyckades Optimize placera ut endast 18 vindturbiner medan motsva-
rande antal för Produktionsmodellen var 28 stycken, vilket innebar en 43%
högre total produktion. För upplösningen 50m var motsvarande antal vind-
turbiner 20 respektive 30 stycken, vilket gav en 38% högre total produktion.
De av Produktionsmodellen utplacerade antal vindturbiner motsvarar ock-
så de maximala antalen för respektive upplösning. Om antalet vindturbiner
begränsades i Produktionsmodellen till de 18 respektive 20 stycken som Op-
timize lyckades placera ut gav denna en total produktion som var ungefär en
procent högre än Optimize i problemet med upplösningen 100m och ungefär
lika i problemet med upplösning 50m. I det senare problemet begränsades
Produktionsmodellen av programvarans tilldelade minne till att inte kunna
inkludera vakeffekter på längre avstånd än 900m. Trots att Optimize och facit
PARK kombinerar multipla vakeffekter annorlunda än Produktionsmodellen
gav denna alltså lösningar med lika bra eller bättre total produktion.
v
kommer än bättre i framtiden kunna, skapa layouter som medför en bättre
avkastning från vindkraftparker.
vi
Contents
Contents vii
1 Introduction 1
vii
2.2.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Roads . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
viii
4.3.1 Foundation costs . . . . . . . . . . . . . . . . . . . . . 42
4.3.4 Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Bibliography 93
ix
x
Acknowledgements
For the mathematical input and discussions, as well for the extensive reading
and improvement of this theses, I would like to thank my supervisor and
examiner, Docent Ann-Brith Strömberg at the department of Mathematical
Sciences at Chalmers University of Technology.
Patrik Fagerfjäll
xi
CHAPTER 1. INTRODUCTION
Chapter 1
Introduction
1
CHAPTER 1. INTRODUCTION
This thesis describes the work of developing models for addressing the opti-
mization of production and cost of the wind farm layout optimization prob-
lem. It relies on existing methods and models for calculating wind resources
in a geographic area, production, wakes, sound propagation and so on, and
is a method for combining these models to optimize the locations of wind
turbines and the road and cable networks. The thesis does not evaluate
physical models or other means of optimization. It will through verification
examples compare one of the models developed to the corresponding module
in the commercial software WindPRO, however, no extensive quantitative
study is undertaken. The aim of the thesis is to develop models and to con-
clude whether it is possible to increase production and profits, rather than
to indicate by how much these may be increased.
The thesis starts by listing many important subproblems of the wind farm
layout design problem. It then gives an introduction to the modeling tool
used in this thesis, mixed integer linear programming. Following these initial
chapters, the two models - the Wind Farm Layout Production Optimization
model and the Wind Farm Layout Infrastructure Optimization model - ca-
pable of solving the problems stated are built, some performance results are
listed and the Production model is verified and compared with commercial
software in a real life example. The application of the models in a real layout
processes is the main focus of the Discussion section.
As of today much of the wind farm layout process is done manually, partly
since there does not seem to exist any good optimization tools (one commer-
cial optimization tool will be presented in this thesis). Care should be taken
so that the iterative work flow of a design process is not restricted too much.
However, due to the enormous amount of information that must be handled
when searching for the best wind turbine locations, there is a lot of room for
computer aided optimization. Although the uncertainties in the models used
for modeling, for example, wind flows and the spread of a wake, need not be
small [22], this is the information that financing decisions are currently based
upon. Thus, increasing the modeled total profit is of great value. Since this
2
CHAPTER 1. INTRODUCTION
profit is highly dependent on the power production, and thus the positioning
of the wind turbines, an effort creating tools that yield a higher production,
even if merely a percent higher, would be well worth.
3
CHAPTER 1. INTRODUCTION
4
CHAPTER 2. WIND FARM LAYOUT
Chapter 2
5
CHAPTER 2. WIND FARM LAYOUT
The first and major part of the wind farm layout optimization problem is to
find the optimum wind turbine locations. This section will define important
factors affecting permitted locations and wind farm production.
Many aspects are taken into account when finding and deciding on project
site area. These aspects include finding a windy and accessible piece of land,
not covered by, for example, wetlands or any type of nature reserve. More-
over, there could be considerations regarding the site area posing as a habitat
to endangered species or being the nesting grounds for birds. Restrictions re-
garding minimum distance from, for example, existing roads could also apply
and should be accounted for when defining the project area. In this thesis it
is assumed this area is defined with restricted areas removed.
6
CHAPTER 2. WIND FARM LAYOUT
18000
60
16000
50
12000 40
10000
30
8000
6000 20
4000
10
2000
0
0
0
5
10
15
20
25
30
0
2
4
6
8
10
12
14
16
Wind
speed
Shift
in
wind
speed
[%]
Wind speed distribution and directions are calculated for all of the site area.
This calculation is based on wind measurements made over a longer time,
preferably from a hub height mast at the project site, which are corrected for
deviance from yearly averages. Photos showing two means of measuring wind
characteristics, a mast and a sodar, are shown in Figure 2.2. Softwares used
in the modeling could be based on computational fluid dynamics (CFD) or
some terrain flow model, and incorporate, for example, elevation, roughness
of the ground surface for a large area surrounding the site [18, 8]. The
output of this software and the input to the thesis models will be Weibull
distribution parameters in a number of angular sectors for all points in a
grid covering the project area and surroundings. The Weibull distribution
has been shown to represent wind speed frequencies well and is commonly
used [24]. A Weibull distribution along with an example of a power curve and
the resulting production distribution is shown in Figure 2.3. An illustrative
representation of the Weibull parameters and wind direction frequency for a
number of sectors is the wind rose, seen in Figure 2.4.
Wind turbines cannot be placed too close to each other. The most obvious
distance is two rotor radii apart to prevent the rotor blades from collid-
ing. There are however usually other requirements or demands on minimum
7
CHAPTER 2. WIND FARM LAYOUT
Figure 2.2: Means for measuring wind speed. Picture (a) shows an 80m
high mast fitted with anemometers whereas in picture (b) a sodar, measur-
ing wind speed at different heights by observing mainly the doppler shift
of returning aucustic impulses, is seen. Both pictures are taken in Sourva,
Sweden.
14
2500
120
Probability
Density
Function
12
100
2000
Power
output
[kW]
10
80
8
1500
60
6
1000
40
4
500
20
2
0
0
0
0
10
20
30
0
10
20
30
0
10
20
30
Wind
speed
[m/s]
Wind
speed
[m/s]
Wind
speed
[m/s]
Figure 2.3: Weibull distribution, a power curve example and the re-
sulting production distribution. The latter is the product of the Weibull
distribution and the power curve at each wind speed.
8
CHAPTER 2. WIND FARM LAYOUT
9 5 9 5 9 5
8
6
8
6
8
6
7
7
7
Figure 2.4: Wind roses, showing frequency and Weibull A and k parameters
for twelve 30 degree sectors, where north is denoted by 1.
Wind passing through a wind turbine will decrease in speed, not to recover
fully for quite some distance. Thus, two wind turbines standing not too
far apart will have a negative effect on each others production, referred to
as wake effects or array losses. This loss of wind speed could be calculated
through a number of models and the optimal positioning of wind turbines will
depend on this effect. These wake models are often based on single turbine
wakes, and the method for combining several wakes varies. One common
way is to sum the kinetic energy deficits, that is the sum of the square of the
velocity deficits. This is said to in some cases have a better resemblance with
reality than just linear superposition, although not generally accepted why
[14, 22, 9]. This thesis will however, due to the linear programming approach,
make use of linear superposition and the differences between these methods
will be addressed later on. For the purpose of manually positioning wind
9
CHAPTER 2. WIND FARM LAYOUT
There are regulations stating maximum sound levels originating from wind
turbines at nearby houses or other sensitive areas. Through the use of models
calculating sound levels around a given source, this relates to a minimum
possible distance between wind turbines and these areas. The regulations
could be formulated in a number of ways, as, for example, "must not exceed
an equivalent sound level of 40 dB(A) at nearby houses, and superposition
of sound pressure from multiple wind turbines is made as if the sensitive
area is downwind of all wind turbines simultaneously", partly taken from
the Swedish Naturvårdsverket [1]. The minimum distance will however vary
depending on the number of wind turbines positioned. Two wind turbines of
the same sort in about the same spot will yield double the sound pressure,
equivalent to a 3dB sound level rise, in a nearby point. This effectively
increases the minimum distance. Sound pressure and sound level relates as
SL = 10 log10 (SP).
The choice of wind turbines will affect the production in a location, the
magnitude of the wake effects and the sound levels. In many wind turbines
there is also an option to decrease the latter at the cost of also decreasing
power production.
It is not only the available amount of wind energy in each point, and thus the
possible power output, that will decide the optimal wind turbine locations
and return on investment. The cost of constructing a wind farm could vary
depending on the amount of work and material needed for foundations as
10
CHAPTER 2. WIND FARM LAYOUT
well as roads and cables connecting the wind turbines, transformer station
and the external power grid. The problem of optimizing this is in contrast to
the above aspects of maximizing production a problem of minimizing costs.
Relative to the cost of wind turbines the costs of roads and cables are rather
small and it is common these costs will not be considered until after the wind
turbine locations have been chosen [24]. If locations are already decided, a
model optimizing the costs could be formulated and used straight away. If,
however, these costs are to be considered when finding the optimal wind
turbine locations, other approaches are needed since the two problems are
optimizing different quantities. One could optimize over both production
and costs by letting one of them be fixed at a certain level, and do this over
and over again using different levels to create a type of trade-off curve, in
economics often referred to as a Pareto efficient frontier. Another approach
is to bridge the two problems by defining a revenue on electricity production,
thus optimizing revenue minus costs over a given time period. This latter
approach will be implemented for the purpose of testing problem complexity.
2.2.1 Foundations
Wind turbines should not be built where the terrain, for example, is too
steep, too rough or consists of too unstable ground for it to be economically
beneficial or even possible to use as a location. To deal with this, the cost of
foundation must be priced correctly in different areas. One should also add
the cost of each wind turbine, though this cost will generally not vary due
to the choice of location. If using multiple types of wind turbines, the costs
of these and possibly foundation costs could vary depending on which type
is chosen. Figure 2.5 shows pictures from the construction of a small wind
farm.
2.2.2 Cables
11
CHAPTER 2. WIND FARM LAYOUT
Figure 2.5: The construction of a wind turbine, with hub height 105m.
Both pictures are taken in Simmatorp, Sweden.
Several cables originating from wind turbines could be combined into high
capacity cables, as in the branches of a tree, with additional costs for con-
nector stations if this connection is not performed at another wind turbine.
If defining a power flow from the wind turbines to the transformer station,
flow can be combined but not split.
2.2.3 Roads
Roads connecting the wind turbines and the outside world are needed for
transporting the wind turbines to their locations and performing mainte-
12
CHAPTER 2. WIND FARM LAYOUT
nance, where the former is the need defining what roads can be built or not.
Wind turbine transports are heavy, with the generator as the heaviest part,
or really long, as when transporting the blades (as seen in Figure 2.6). This
calls for robust, low-incline roads, which, for example, could result in it not
being possible to build on top of a hill even though wind energy levels might
be high and foundation costs low.
Figure 2.6: Wind turbine blade transport. Each blade is 45m long. The
picture is taken in Simmatorp, Sweden.
A large wind farm would usually connect to the external power grid through
a transformer station which, just as in placing a wind turbine, will result
in foundation as well as road and cable costs along with the cost of the
transformer station itself. Its optimal placement is dependent on the cost of
the high voltage export cable needed for connecting to the external power
grid compared to the cables needed for connecting the wind turbines. The
construction cost of the transformer station can also vary depending on its
location within the farm. Power losses in the transformer will depend on the
current, and thus the number of wind turbines, but not on the location of
the wind turbines within the farm. It is possible to allow for different sizes
of transformer stations if the number of wind turbines to be placed is not
decided prior to optimization.
13
CHAPTER 2. WIND FARM LAYOUT
14
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
Chapter 3
15
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
This section gives a short introduction to the field of mixed integer linear
programming used in developing the models of the thesis. Examples and
basic theory of linear programs, integer linear programs and the definition
of graphs and trees are presented, along with a final section defining the
notations used in this thesis.
The word linear in linear programming (LP) refers to the variables of a prob-
lem being combined in a linear way, that is adding or subtracting variables
but not, for example, multiplying them or using them as exponents. Most
physical problems, or other types of problems for that matter, cannot be
expressed in a linear way, but those who can can be solved in more efficient
ways. The concept of linear optimization will here be explained by the use
of an example.
Let us say we have discovered that the citizens of New York have great
cravings for Swedish Knäckebröd and Messmör, a form of crispbread and
soft whey butter. We see a great business opportunity and decide to bring a
cart and a bag full of these goods across the Atlantic to pursue a career as
Knäckebröd and Messmör vendor. On the flight to America the cart will take
up all of our checked baggage allowance, leaving us with just the carry-on
baggage for transporting our goods. If the economic and physical properties
of Knäckebröd and Messmör are according to Table 3.1 and the carry-on
baggage is restricted to 8 kg and 60 litres respectively, what quantities of
the two products should we bring to maximize our profit in the land of
opportunities? For the moment, fractions of packages are allowed to be
bought and sold and we are able to pack the goods real tight in our baggage.
16
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
The total potential profit, which is what we want to maximize, can be ex-
pressed as the linear function of packages of Knäckebröd, K, and Messmör,
M , called the objective function, to
17
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
18
16
14
weight = 8kg
12
knäckebröd,
k
10
8 volume = 60l
6
4
2
0
0
2
4
6
8
10
12
messmör,
m
18
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
18
16
14
weight = 8kg
knäckebröd,
k
12
10
8 volume = 60l
6
4
2
0
0
2
4
6
8
10
12
messmör,
m
To get a hint of why integer problems take a long time to solve, one could
compare to the complete enumeration of all possible combinations. If hav-
ing a set of n binary variables, xi ∈ {0, 1}, where i = 1, ..., n, the number
of possible ways to combine these variable values is 2n . With a computer
capable of computing and checking feasibility of one billion, 109 , of these
combinations per second, the time needed for different number of variables
is shown in Table 3.2. If the problem is that of connecting n wind turbines
by roads, similar to the formulation derived in Section 4.3.2, the number of
integer variables correspond to n + n2 , and the complete enumeration time,
using the same computer as above, is also seen in the table.
19
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
n time, 2n
20 0.001 seconds
50 313 hours
80 38 years
2
n time, 2n+n
5 1 second
6 1.2 hours
7 2.3 years
8 150 thousand years
20
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
The principle of the branch and bound method is illustrated in Figure 3.4
for the integer problem example of exporting Knäckebröd and Messmör, and
is described below.
Continuing the example problem 3.1, we may sell fractions of packages, but
when buying our goods in Sweden we will have to buy full packages. This
translates into requiring the variables representing Knäckebröd and Messmör
being integer, that is K M ∈ N, where N = 0, 1, 2 ... which is altered in the
mathematical formulation to
Now, to solve this problem the branch and bound algorithm is used. The
procedure is hereby described and is illustrated in Figure 3.4. We start
off with the optimal solution to problem 3.1, (K, M ) = (7.6, 5.3). Now,
we choose a fractional variable, in this case K, and branch in to the two
continuous subproblems with the additional constraints K ≤ 7 and K ≥ 8,
respectively. When solving these continuous problems we notice an integer
solution, (8, 0) = 160, found in the "K ≥ 8"-branch. This is a solution to
the integer program 3.2, but we do not know whether it is optimal. The
MIP-gap tells that the value of the real optimum cannot be more than 75%
better than this integer solution. This branch can now be pruned and the
objective value of 160 is used as a lower bound on the optimal objective value.
If another branch yields an upper bound (a LP-relaxation value) lower than
this, that branch can be pruned. Continue by branching on the fractional
value of M in the leftmost node in level 1, yielding an even better integer
solution, (7, 5) = 265, and another fractional solution, respectively. From
21
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
18
16
14
weight = 8kg
12
knäckebröd,
k
10
8 volume = 60l
6
4
2
0
0
2
4
6
8
10
12
messmör,
m
Figure 3.3: The feasible set of solutions to the integer problem. The
dots all represent possible solutions.
this node we branch into K ≤ 6 and K ≥ 7, but due to the earlier constraint
of K ≤ 7 the only option for the rightmost branch is K = 7. Now, the
configuration of M ≥ 6 and K = 7 is infeasible and searching along that
branch is stopped. We carry on, find an even better integer solution, (6,6),
with an objective value 270, but still does not know whether it is optimal.
The next split however yields one infeasible subproblem and one for which
the upper bound is less than the objective of the best integer solution found,
and we have proven the solution (K, M ) = (6, 6) to be optimal.
The feasible set of solutions to this problem is once again, along with this
computed optimum, illustrated in Figure 3.5.
22
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
(7.6, 5.3)
Level 0 MIP gap
283
K≤7 K≥8
(7, 5) (6.4, 6) €
278 278
Level 2 265 −1 = 4.9%
K≤6 K=7
265
€ 271
(6, 6) (4.8, 7) −1 = 0.4%
Level 4 270 271
270
K≤4 K≥5
€
(4, 7.5)
Level 5 infeasable
267.5
Figure 3.4: Branch and bound solution tree, for the problem of exporting
Knäckebröd, K, and Messmör, M.
23
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
18
16
14
weight = 8kg
12
knäckebröd,
k
10
8 volume = 60l
6
4
2
0
0
2
4
6
8
10
12
messmör,
m
2
1 3
4 5
y6,1
6 7
There are numerous optimization problems formulated using graphs, as, for
example, the traveling salesperson problem, where one needs to find the
shortest path for a salesperson visiting every city in a set but never visiting
24
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
the same city twice, or as similar to the problems of this thesis, to find low
cost tree structures spanning the nodes. A minimum spanning tree, for ex-
ample, is to find the shortest, or cheapest, combination of edges that connect
each node to at least one of the other nodes, creating a structure of edges sim-
ilar to the branches of a tree. For many of the classic graph theory problems,
heuristic solution algorithms have been developed. However, when altering
such a problem or creating a somewhat different type of graph problem, as
in this thesis, these heuristic algorithms need not be applicable. Finding effi-
cient algorithms producing at least close to optimal solutions using heuristics
developed for a new problem could prove to be hard.
25
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
26
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Chapter 4
27
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
28
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
X
maximize p i xi ,
i∈V
X
subject to xi ≤ n,
i∈V
xi ∈ {0, 1}, i ∈ V.
To prevent wind turbines from being placed too close to each other, con-
straints regarding the minimum separation distance between wind turbines
are formulated. This is done by stating that none of the other nodes within
a given range from an active node is allowed to be active. This minimum
separation distance is defined as a factor α times the rotor radius, R, when
modeling a circular minimum separation zone, but it can also be expressed
by for instance an ellipse with it’s major axis aligned with the primal wind
direction.
Setting the sum of all nodes within the minimum separation distance αR of
node i less than or equal to one will not do the trick since the constraint
has to hold also for a node where no turbine is placed. At this inactive
node it is perfectly alright for more than one node within the distance αR
to be active without them being less than αR meters apart, as illustrated in
Figure 4.2. In a continuous representation of possible locations this would
lead to a formulation where a maximum of five nodes strictly less than the
minimum separation distance from i would be allowed when node i is not
active, and one otherwise. Six active nodes around an active or non active
node is possible only when the distances from these to i is larger than or
equal to αR, shown in Figure 4.1. The mathematical representation of this
29
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
constraint will be the sum of five times the value of node xi plus the values
of the surrounding nodes xj , where j ∈ Qi , should be less than or equal to
five. The set Qi is defined as all the nodes j with distance from i less than
αR, except for the node i itself. |Eij | denotes the distance in meters between
the nodes i and j. The constraints read
X
5xi + xj ≤ 5, i ∈ V,
j∈Qi
where Qi = j ∈ V |Eij | < αR \ {i} , i ∈ V . The solution of this problem
is however slow and a different approach is therefore adopted. Here, the sum
of active nodes within a certain radius around a node is set to be less than
or equal to one. The radius used is of half the minimum separation distance
and now the program rely on the points in between each pair of nodes to
set the minimum separation distances, as illustrated in Figure 4.4. At most
one wind turbine may be placed within each of these circles. The problem
becomes one of packing circles as dense as possible, around active and non
active nodes, as seen in Figure 4.3. To allow for additional temporary nodes,
used only for these constraints but not as possible wind turbine locations, a
new set of nodes and edges, Ḡ = (V̄ , Ē) where G ⊆ Ḡ, is introduced. The
new constraint, which now will be used in the formulating of this Production
model, is given by
X
xj ≤ 1, i ∈ V̄ , Qi = j ∈ V̄ |Ēij | < αR/2 .
j∈Qi
The constraint says that the sum of the active nodes in a set Qi may not be
greater than one, where Qi consists of all nodes j on distance less than αR/2
from node i (including i itself). |Ēij | denotes the distance in meters between
the nodes i and j.
30
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Figure 4.1: Six wind turbines in an area less than the minimum
separation distance of a node is not possible. The tightest placement
of the six are instead on the minimum separation distance.
31
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Figure 4.4: A maximum of one active node in the circle area less
than half the minimum separation distance of a node, forces two
active nodes to be separated by at least the minimum separation distance.
32
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
To allow for close to hexagonal packaging, a grid size of the minimum sepa-
ration distance divided by eight or multiples of eight is recommended. On a
bounded area, for which a hexagonal or a regularly spaced packing is optimal,
optimization on a discrete grid may yield a solution close to the continuous
optimum. Examples of this are shown in Figure 4.5, where the proven con-
tinuous optima are obtained from [11] and [10].
To account for wake effects, that is the loss of production when placing two
wind turbines close to each other (see Section 2.1.4), a new variable wij ∈
{0, 1} is introduced. This variable, if equal to one, represents a production
reduction at node j due to a wind turbine at node i. This constraint equals
the one of the generalized vertex packaging problem [20] and is also used
by Donovan in [4, 5] and Donovan et al. in [6]. The magnitude of the
corresponding production loss is denoted dij MWh and must be calculated for
every pair of nodes since it depends on the wind frequency and distribution
for the specific location and direction of wake. Since the proximity range
constraints of Section 4.2.2 prevent the use of edges shorter than αR, these
can be excluded from the set E, defining the subset Ê ⊆ E. The modified
objective function is to maximize
X X
p i xi − dij wij ,
i∈V (i,j)∈Ê
33
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
1800 1600
1600 1400
1400 1200
1200
1000
1000
800
800
600
600
400
400
200
200
0
0
−200
−200
0 500 1000 1500 0 500 1000 1500
[meter] [meter]
1800
1600
1400
1200
1000
800
600
400
200
−200
0 500 1000 1500
[meter]
(c) 25 circles
Figure 4.5: Circle packing. Placement of 19, 20, and 25 wind turbines,
respectively, in different geometric areas. The inner black dashed line rep-
resents the area required for packing the number of circles in a continuous
manner, whereas the outer dashed line is the area required using the for-
mulated linear program. The two dashed lines differ little or nothing in the
three examples. The solid line is the project area used by the linear program.
34
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
The variables wij ∈ Ê may take only binary values, but due to the formula-
tion of the constraint relating two nodes and their edges this requirement will
be fulfilled in an optimal solution alsoP without integrality requirements on
wij . The sum of the losses from wakes, (i,j)∈Ê dij wij , is subtracted from the
objective function, and the wake constraint forces wij to 1 if both xi = xj = 1
(i.e., the nodes i and j are active), otherwise it will take the value of 0.
The formulation of wake effects between two possible locations using linear
relations is not restricted to any particular wake model. However, due to
linearity the combination of multiple wake effects on a node equal the super-
position of all wake effects from active locations on this node. If non-linear
superposition is preferred, approximations by linearizations are addressed in
Section 4.2.7.
To deal correctly with maximum allowed sound levels one cannot simply de-
fine a minimum allowed distance to wind turbines from a noise sensitive area,
denoted by an index s ∈ S̄, corresponding to a certain sound level. This is
because the sound pressure from two wind turbines will be superposed and
the minimum allowed distance thus increased. The constraint regarding max-
imum sound levels is formulated as the superposition of the individual sound
pressures ais in s from active nodes i, which must not exceed a maximum
allowed value, amax . When using multiple types of wind turbines, their differ-
ent sound characteristics must also be implemented in this constraint. Sound
pressure (SP) and sound level (SL) relate as SL = 10 log10 (SP). The resulting
constraint is given by
P
i∈V ais xi ≤ amax , s ∈ S̄.
35
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Combining the formulations in Sections 4.2.1, 4.2.2, 4.2.4 and 4.2.5 yields
the complete Production model, which is here described and presented.
The Production model optimizes the total production when placing wind
turbines in a project area. It can place at maximum a given number, n, or as
many wind turbines as possible. The model respects the minimum separation
distance between wind turbines, αR, and that the sum of the individual sound
pressures, ais , originating from the wind turbines, is held below the maximum
level, amax , for any noise sensitive area, as . The production, pi , in each
possible location, and the production loss, dij , between wind turbines i and j
due to wake effects, are calculated from any production and wake effect model
(see Sections 2.1.2 and 2.1.4), respectively. The graph G = (V, E) consist of
all the nodes V and the edges E connecting each pair of nodes. Ĝ = (V̂ , Ê) is
the subgraph of G with edges shorter than the minimum separation distance,
αR, removed, and Ḡ = (V̄ , Ē) is the graph with twice the grid resolution of
G, used for implementing the minimum separation distances. The complete
Production model is to
X X
maximize p i xi − dij wij ,
i∈V
X(i,j)∈Ê
subject to xi ≤ n,
i∈V
X
xj ≤ 1, Qi = {j ∈ V̄ |Ēij | < αR/2}, i ∈ V̄ ,
j∈Qi
xj − wij ≤ 1,
xi + X (i, j) ∈ Ê,
ais xi ≤ amax , s ∈ S̄,
i∈V
xi ∈ {0, 1}, i ∈ V,
wij ∈ [0, 1], (i, j) ∈ Ê.
36
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
These suggested formulations or ideas have not, due to time restrictions, been
implemented in this thesis. They seem however well worth to investigate in
future work.
X X
xk + xk − 1 ≤ wij , (i, j) ∈ Eijs ,
k∈Si k∈Sj
where Si ⊆ V , i = 1, ..., |S|, is the subset of the nodes within square sector
i. The set E s are composed by the edges connecting each pair of sectors.
37
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
14
12
Cumula&ve
distribu&on
[%]
10
8
6
4
2
0
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Rela&ve
movement
[%]
38
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
0,14
0,12
Cumula&ve
distribu&on
[%]
0,1
0,08
0,06
0,04
0,02
0
0
1
2
3
4
5
6
7
8
9
10
Angular
shi8
[deg]
Figure 4.8: Angular shift, showing the cumulative distribution of the shift
in bearing between two randomly placed wind turbines when shifting their
locations to the sector centers. One wind turbine is placed in the square sector
with centre in the point (x, y) = (0, 0), the other within x, y ∈ [−2000, 2000],
respecting the minimum separation distance of 400m. The sector side length
is 100m.
X
wij ≥ (K − 1)xi , i ∈ V̄ ,
j∈E
X
wij ≥ (K − 1)xj , j ∈ V̄ .
i∈E
These constraints do, however, only force one of the nodes to which the wake
39
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
is connecting being active, and the use of small wakes to and from non-active
nodes at a far distance is possible. Additional constraints guarantee that the
nodes that an active wake is connecting to are also active, according to
wij ≤ xi , i, j ∈ V̄ ,
wij ≤ xj , i, j ∈ V̄ .
These four inequalities imply that the number of active wakesPrelatesP cor-
rectly to the minimum allowed number of active nodes, i.e., i j wij =
K(K − 1). This does, however, not force the LP-relaxation to take on in-
teger values. It is, for example, possible to split two quite separated wind
turbines into two times two 0.5-valued ones next to each other. This allows
for two times 0.5 valued wakes leaving each half wind turbine, resulting in
a total wake value of two instead of one between the original wind turbine
locations. The LP relaxation will by this splitting behavior yield large sums
of wake variable values between distant nodes, and less between nodes close
to each other.
40
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
XXX
maximize pni,θ Ki,θ
n
,
i∈V
Xθ∈Θ n∈N X
n (4.1)
subject to Ki,θ ≤ wki , i ∈ V, θ ∈ Θ,
n∈N k∈S(i,θ)
n
Ki,θ ∈ {0, 1}.
XXX
maximize pnl,θ Kl,θ
n
,
Xθ∈Θ n∈N
l∈L X
n (4.2)
subject to Kl,θ ≤ xi , l ∈ L, θ ∈ Θ,
n∈N i∈S(l,θ)
n
Kl,θ ∈ {0, 1}.
41
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Another optimization area of interest concerns the costs associated with the
constructions of foundations as well as roads and cables necessary for the wind
farm. In these formulations the locations of the wind turbines are considered
fixed. However, the simultaneous solution of the Production model and the
Infrastructure model, not requiring the locations to be fixed, is possible but
seems to be a rather complex problem.
X
maximize − ci x i ,
i∈V
subject to xi ∈ {0, 1}, i ∈ V,
which, if the locations of the wind turbines are fixed by the prior solution of
the Production model, can only take on one value.
When connecting the wind turbine locations with roads and power lines, a
spanning tree structure, as defined in Chapter 3.3, is to be constructed. An
example of this is connecting the nodes of Figure 4.9 as shown in Figure 4.10,
a minimum spanning tree solution obtained by the model described below.
This example will be addressed again when discussing cables in Section 4.3.4.
Each edge is associated with a certain cost, dij defined by a number of fac-
tors such as length, topography and such. A linear programming formulation
which will minimize the sum of the edge costs is sought. However, the formu-
lation should also allow for, but not require, the tree to use paths involving
nodes that are not active as locations of wind turbines. Such a formulation
42
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0 1000 2000 3000 4000 5000
19 10
4500 10 22
1
2
16 11 22 23
11 9 3
4000 9 8 23 16
8 13 7 2 23
1 13 9 3
3500 4 9 6 4 21
14 21 2 1
24 13
2
3000 20 24 2
1 6 17
15
2500 18 17 17 7
1 17
7 12
2000 18
1500
12 14
19
26 25 14 26
1000 23 22
25 3
24 15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.10: A minimum spanning tree considering only the layout of the
edges and not minding that they are directed, connecting the nodes of Figure
4.9. There are no limits on the flow through the edges. The upper two figures
of each triple of numbers state which two nodes the edge is connecting and
the bottom figure state the flow through the edge.
43
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
could allow for a shorter tree (if defining an euclidean tree) or ease the pro-
cess of formulating paths around obstacles (Figure 4.11), and results in a so
called Steiner tree problem [3, 2]. Since the problem is that of electricity
flow, at some central point of the tree the flow should be passed on to an
external grid.
Figure 4.11: Shorter paths are possible, when using other nodes than
just the active ones (Steiner tree formulation). The left path is 5% longer
than the right one.
X X
maximize (ri − ci )xi − dij zij , (4.3)
i∈V (i,j)∈Et
To deal with the actual creation of the tree structure, and to prevent the
occurrence of subtours, a new continuous variable yij ∈ [0, f ], representing
the flow of electricity in the tree, is introduced. Here, f ∈ [0, n], where
n is the maximum allowed number of nodes connected at, or flow in, edge
(i, j). Each active node creates one unit of flow and flow entering a node
must also leave the node. The accumulated flow at the connection node (the
external power grid), denoted CN ∈ Vt , must equal at least the number
44
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
of active nodes minus one. We denote the flow on edge (i, j) by yij . This
connection node constraint is not really necessary, since the flow must be
exported somewhere and the connection node is the only option, but it was
discovered having a slight effect on reducing the time of finding a feasible
solution. The constraints read
X X
yij + xj − yjk = 0, j ∈ Vt \{CN },
ij∈Et jk∈Et X
(4.4)
P
i∈Vt yi,CN ≥ xi − 1,
i∈V
yij ∈ [0, fij ], (i, j) ∈ Et .
Now, yij is the number of wind turbines connected on the branch from node
i, including i, and it’s flow can be limited either individually for each edge
(i, j) or for all (i, j) ∈ Et . The binary variable defining the activation of
an edge is however zij and to induce the activation and cost of an edge we
connect these two by replacing the last line of (4.4) by the constraints
Using the formulation illustrated in (4.4) with the modification of (4.5), iso-
lated subtours are not possible due to the accumulative property of the vari-
ables yij , illustrated in example Figure 4.12. There exists feasible solutions
with subtours connected to paths leading to the connection node, but it is
then cheaper to break up the subtour, as illustrated in Figure 4.13.
3+1
0+1
1+1
2+
1
Figure 4.12: Isolated subtours, are not possible since flow is added in each
active node and the flow along an edge can take on only one value.
The basic tree formulation can then be written as the linear program to
45
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
3 3
1
3 2
Figure 4.13: Non-isolated subtours, are possible, but breaking the subtour
will yield a shorter path.
X X
maximize ri xi − dij zij ,
i∈V (i,j)∈Et
X X
subject to yij + xj − yjk = 0, j ∈ Vt \{CN },
i:(i,j)∈Et k:(j,k)∈Et
X X
yi,CN ≥ xi − 1,
i∈Vt i∈V
0 ≤ yij ≤ fij zij , (i, j) ∈ Et ,
xi ∈ {0, 1}, i ∈ V,
zij ∈ {0, 1}, (i, j) ∈ Et .
relates the flow through and activation of an edge in the opposite way that
(4.5) does. The need for at least one active edge variable, zij , per active node
is expressed by
X
xi − zij ≤ 0, i ∈ V.
j:(i,j)∈Et
46
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
A flow leaving a node requires either flow entering the node or the node itself
being active, which is formulated by the constraint
X X
zij + xj − zjk ≥ 0, j ∈ Vt \{CN }.
i:(i,j)∈Et k:(j,k)∈Et
Moreover, forcing flow in only one direction per edge is beneficial and can be
expressed by
It is possible to limit the number of edges per node in a grid to only the
shortest edge in, say, sixteen directions, resulting in a bit more costly span-
ning tree but reducing the number of variables zij and yij to 16|Vt | instead
of |Vt |2 each. If roads are to be connected along an existing road instead
of at a single defined point, nodes with corresponding edges could be added
along this road and the cost of these edges set to zero, as in the example
illustration of Figure 4.15. Since roads and cables typically follow the same
paths – due to the need of a road to place a cable – the cost of a road is
merely an additional cost added to the cable cost, defined in Section 4.3.4,
and no additional road formulations to the Steiner tree are thus given here.
Two cables along the same path only requires one road to be built, which
is solved in the same manner as for a single cable by defining double cables
as a cable type in the mathematical formulation. This implemented model
will be able to handle double cables when flows are parallel. However, in
reality two cables with opposite flows will also be beneficial in terms of road
costs, as per the rightmost illustration in Figure 4.14. Another formulation
addressing this issue, where separate road variables are introduced, is given
in Section 4.3.7, however not implemented.
47
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
1 1
1
1 2
2
2
Figure 4.15: Using existing roads may reduce the cost of the path. Costs
are given in the diamond symbols. The cable path following the existing road
is longer but cheaper than the straight path.
48
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
XX
maximize − Cki Rki ,
X i∈Vt k X
subject to zij − Rki ≥ 0, i ∈ Vt ,
j∈Vt X k∈K
X
Rki ≥ 1,
i∈Vt k∈K
Rki ∈ {0, 1}, i ∈ Vt , k ∈ K.
4.3.4 Cables
49
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
different costs, is shown in Figure 4.18. Whether the solution time will ben-
efit from an additional constraint restricting the use of only one zijl per edge
(i, j), which in practice will always be the case, has not been investigated.
19 10
4500 10 22
1
2
11 9
4000 1 23 16
8 13 1
1
3500 9 6 4 21
2 21 2 1
16 7 2
3000 20 24 13 17 2
1 2
22 26
24 18 3
2500 2
6 26 2 26
3 3
2000 17 26
26 7 26
18
33 3
1500 12 26
1
25 26 14 26
1000 2 3
3 25
126 1
15 14
500 25 2 5 15
1
0
0 1000 2000 3000 4000 5000
Figure 4.16: An upper limit of three wind turbines per edge. With a
cable cost of 0.6 MSEK/km, and export cable cost of 3 MSEK/km, the total
cost of connecting the nodes of Figure 4.9 is 22.7 MSEK. This configuration
was found after about 25 s with a MIP-gap of 4% and was proven optimal
after 243 s. The cable connecting to the external power grid in point (0,0) is
the high capacity export cable originating from the transformer station.
Now, the transition between two cable types will occur either when passing a
wind turbine, thus shifting the capacity demand, or when combining the flow
from two different locations into one cable. The former case as well as a split
at a wind turbine location is considered to come without extra costs, since the
necessary equipment is already present. However, when the cables are to be
combined at a location other than at a wind turbine, the cost of the required
substation has to be accounted for. This is formulated by introducing a
binary substation variable, sj , where j ∈ Vt , with the associated cost gj ,
which should be forced to a value of one if flows are combined outside of
wind turbine locations. If the combination occurs at the location of a wind
turbine, sj = 0. This is achieved by requiring either sj or xj to take on a
value of one if more than one flow enters the node j. The flow combination
program is thus to
50
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
19 10
4500 10 22
1
2
11 9 16 11
4000 2 23 16 22 4
3 3
8 13 1
1
3500 9 6
4 4 21
13 24 21 2 4
5
3000 20 24 2 2 7
1 6 17
5 6
2500 17 18
6
2000
24 3
4 7 3
18 3
1500 7 7
12 14
1
26 25 14 26
1000 5 4
25 3
6
15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.17: An upper limit of seven wind turbines per edge. With a
cable cost of 1 MSEK/km, and export cable cost of 3 MSEK/km, the total
cost of connecting the nodes of Figure 4.9 is 24.3 MSEK. This configuration
was found and proven optimal after 11 s. The cable connecting to the external
power grid in point (0,0) is the high capacity export cable originating from
the transformer station.
X
maximize − gj sj ,
X j∈Vt
subject to zij − 1 − M (xj + sj ) ≤ 0, j ∈ Vt , (4.6)
i∈Vt
sj ∈ {0, 1}, j ∈ Vt ,
Flows can be combined, but are not permitted to be split, which calls for
additional constraints. These constraints, limiting the number of flows out
of a node i, are defined as
X
zij ≤ 1, i ∈ Vt .
j∈Vt
51
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
19 10
4500 10 22
1
2
22 23
11 9 3
4000 1 23 16
8 13 4
1 9 13
3500 3 6 9
1 4 21
13 24 21 2 1
16 17 2
3000 20 24 5 5 2 7
1 3
2500 17 18
6
7 12
2000 4
24 3
7
18 3
1500 7 12 26
5
26 25
1000
25 3 14 6
3
7 3
15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.18: Multiple cable variables with upper limits of three and
seven connected wind turbines, respectively. Using the prices from
Figures 4.16 and 4.17 the total cost of connecting the nodes of Figure 4.9
is 20.4 MSEK. This configuration was found after about 50 s with a MIP-
gap of 6% and was proven optimal after 47 min. The cable connecting to the
external power grid in point (0,0) is the high capacity export cable originating
from the transformer station.
52
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
X
maximize − ctr
i ti ,
i ∈Vt \{CN }
subject to zi,CN ≤ ti , i ∈ Vt ,
ti ∈ [0, 1].
The Infrastructure model is that of connecting the wind turbines by roads and
cables, and of locally varying foundation costs. If the locations of the wind
turbines are fixed, the location variables xi could be set to their defined values
and no revenue, ri , would be necessary. For the purpose of combining the
53
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
two models, however, this revenue is included in the objective function. The
graph Gt = (Vt , Et ) contains the nodes and edges used to form the Steiner
tree. The set, V , of possible or fixed locations of wind turbines is a subset
of Vt . The nodes of Vt can be many more than the number of wind turbines,
making cheaper connections possible, and includes the point of connection
to the external grid, xCN . Power flow along an edge is represented by yij
and one unit of flow is added in each active node along a cable. The sum of
the flow entering the connection node must not be less than the number of
active nodes, or this number minus one if wind turbine placement is allowed
in the connection node. Due to this accumulated flow and the need to export
the flow, isolated subtours are prohibited. Several types of cables, yijl , where
l = 1, 2, ..., with different maximum flows, fijl , can be used on an edge. The
cost, dlij , of cable and road type l on edge (i, j) is activated by the binary
variable zijl . Moreover, constraints requiring at least one road Rki with cost
Cki be connected to the road tree are included. There are also constraints
introducing a substation cost gj through the combination variable sj where
flow is combined outside a wind turbine location. Flow splits are prohibited
by another constraint, and the location specific cost of a transformer station,
ctr
i , is introduced at each transformer station location by the variable ti . The
complete Infrastructure model is to
54
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
X X XX X X
maximize r i xi − dij zij − Cki Rki − gj sj − ctr
i ti ,
i∈V (i,j)∈Et i∈Vt k j∈Vt i ∈Vt \{CN }
X X
subject to yij + xj − yjk = 0, j ∈ Vt \{CN },
i:(i,j)∈Et k:(j,k)∈Et
X X
xi − 1 ≤ yi,CN ,
i∈V i∈Vt
0 ≤ yij ≤ fij zij , (i, j) ∈ Et ,
X yij ≥ zij , (i, j) ∈ Et ,
xi − zij ≤ 0, i ∈ V,
j:(i,j)∈Et
X X
zij + xj − zjk ≥ 0, j ∈ Vt \{CN },
i:(i,j)∈Et j:(j,k)∈Et
X X
zij − Rki ≥ 0, i ∈ Vt ,
j∈Vt X k∈K
X
Rki ≥ 1,
X i∈Vt k∈K
zij − 1 − M (xj + sj ) ≤ 0, j ∈ Vt ,
i∈Vt
zi,CN ≤ ti , i ∈ Vt ,
xi ∈ {0, 1}, i ∈ V,
zij ∈ {0, 1}, (i, j) ∈ Et ,
Rki ∈ {0, 1}, i ∈ Vt , k ∈ K
sj ∈ {0, 1}, j ∈ Vt
ti ∈ [0, 1], i ∈ Vt .
If turbine positions are held fixed the number of nodes in the grid defining
the tree structure can be reduced. It would, for example, generally not be of
any use to keep nodes on the outside of the "outermost" wind turbines. Also,
better geographic locations for combining flows, for example, at the centre of
a group of wind turbine locations, can be found for various combinations of
the wind turbines. Thus, the set of nodes can through a bit of smart prepro-
cessing be redefined to a smaller but better placed set of nodes. The grid can
55
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
If the Infrastructure model is combined with the Production model, and thus,
the locations of the wind turbines are not fixed, this type of rearrangement
is not possible. However, a similar approach to that of defining wake effects
between sectors (Section 4.2.7) may be adopted, creating the tree structure
connecting geographical sectors instead of specific points (which are included
in these sectors).
Two or more cables on the same path is in the Infrastructure model formu-
lated by an edge variable with the cost of multiple cables but the cost of
only one road. However, it cannot handle, for example, opposite directions
of flow, which would also lead to only one road having to be built. A sim-
ple solution to this problem, increasing the number of binary variables and
56
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
c
zijc + zji ≤ 2zijr , {j ∈ Vt | j > i}, i ∈ Vt ,
r
zij ∈ {0, 1}.
This solves the problem of the costs not decreasing when using a path as in
the rightmost illustration of Figure 4.14. However, due to the restriction of
not allowing splits of flow, this type of path, where flow enters and leaves
and then enters and leaves the same location again, is still not possible. A
solution to this is to also implement the idea in the next section.
X X
zicm ,jn + zjcn ,im ≤ M zijr , i, j ∈ Vt : j > i,
m={1,l} n={1,l}
where M ≥ l. This shift of nodes and edges requires changing the sets used
in many of the other constraints too.
The problem size increases rather fast by the introduction of these multiple
57
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
nodes, mainly because the number of edge variables yijl and zijl grows by the
square of this multiple. The number of variables needed in different locations
of the project area could however vary, as, for example, shown in Figure 4.20,
which would result in the problem size growing at a slower rate.
Now, this formulation is preferably used when the locations of the wind
turbines are fixed, whence multiple nodes are not needed at these locations,
and the rearrangement of nodes to geographically superior locations, as in
Figure 4.19, can be implemented.
In this formulation, the values added to the flow through the cables of the
58
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Σ
Figure 4.21: Wind farms connected to a high voltage export line.
The farms can be combined in substations and must be connected with a
transformer station somewhere along the export line. The node in which
the incoming flow is accumulated does not have to be where the transformer
station is placed.
Steiner tree are that of the number of wind turbines in the connected parks.
Thus, the variable yij represents the power flow and the value of this flow
in a point on the transmission line equals the total number of connected
wind turbines. Two modifications to the original problem of substations
(4.6) are addressed. First of all, the costs of substations combining flows is
dependent on the magnitude of the entering flows. An additional constraint
is introduced, which forces one of the binary substation capacity variables
Kjp , where p is a capacity in the set Pj , of node j, to take a value of one if the
value of the substation variable, sj , equals 1. The cost, gjp is now associated
with Kjp , where gjp2 ≥ gjp1 if p2 ≥ p1 . The new formulations for substation
costs are to
XX
maximize − gjp Kjp ,
j∈Vt p∈Pj
X X
subject to yij − M (1 − sj ) − pKjp ≤ 0, j ∈ Vt ,
i∈Vt p∈Pj
Kjp ∈ { 0, 1}.
Now, this formulation generate costs when combining cables, thus requiring
substations. To deal with the connection and transformation to a existing
59
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
high voltage export line, a connection which could have only one input cable,
an additional constraint is needed. If the connection could occur at any
number of nodes along this line, this set of nodes, Vel , is subtracted from the
additional substation constraint just formulated and is instead included in
the constraint of the linear program to
XX
maximize − gjp Kjp ,
Xj∈Vel p∈P
Xj
To handle the arbitrary connection point on this export line, the flow gath-
ering connection node of the tree structure is set as one of the nodes one the
export line with flow variables yij and zij from the other nodes in Vel being
without limit and cost respectively. Thus, even though summing the flow in
one specific point, connection to the export line could occur anywhere along
its path.
60
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
61
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
62
CHAPTER 5. TESTS AND RESULTS
Chapter 5
63
CHAPTER 5. TESTS AND RESULTS
This chapter starts off by listing optimization results for a real wind farm
project site, shown in Figure 5.1, and the comparison of this optimization
to the one of the commercial software WindPRO 2.6 Optimize, from now
on referred to as Optimize. Further on, the performance of the Production
and Infrastructure model in terms of speed and MIP-gap is expressed. The
total production of any layout found will be calculated through the software
WindPRO 2.6 PARK, from now on referred to as PARK. For references on
MIP-gap and other mathematical terms, see Chapter 3.
3500
3000
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
Figure 5.1: Project area of the verification examples. The green area is
project area whereas the pink areas are wetlands.
As up to this point, the models have been invariant to how production, wakes
and costs are calculated. For testing purposes however, a choice of these
methods is needed. In the following results, production is calculated using a
power curve for a Vestas V90 2MW wind turbine at 105m hub height, based
on an average of the site atmospheric pressure. The wind energy content is
calculated through WindPRO using the WindATLAS method [8, 18] and ar-
ray losses are expressed through the Jensen/Risø single wake model [13, 14].
The sound pressure level propagation model used is the one recommended
by Naturvårdsverket [1]. Furthermore costs of wind turbines, roads, power
lines and the price of electricity are set at reasonable levels. The models
are implemented as matrices in Matlab and optimization is carried out using
the optimization package Cplex 12.1 with its Matlab application program-
ming interface. The computers used are i686’s with four Dual Core AMD
64
CHAPTER 5. TESTS AND RESULTS
Examples of the layout is shown in Figure 5.3 for the 100m resolution with
and without wake effects considered. Not modeling wake losses does not
resemble reality but can be used as an aid in parts of the optimization pro-
cess. It will, if wind energy content does not vary much or not at all within
the project area, translate into fitting as many wind turbines as possible.
This maximum number is valuable when trying to find the maximum total
production of an area.
The maximum numbers of 28 and 30 wind turbines for the grid resolutions
100m and 50m, respectively, is also found when array losses decrease the
production in tight placements. An example of this is shown for the 100m grid
in Figure 5.3. In this example, the PARK total production is practically the
same no matter whether the wake effects are considered in the optimization
process or not, which is seen in Figure 5.4. A reason, other than the limited
number of ways of fitting 28 wind turbines in the area, for the solution
65
CHAPTER 5. TESTS AND RESULTS
3500
2.00
3000
1.75
2500
1.50
2000
1.25
1500
1.00
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
Figure 5.2: Wind energy content. Scaled to values of one being the
minimum and two the maximum energy content within the project site area.
The energy content is calculated using the WindATLAS method, which uses
topography as well as other inputs.
3500
3500
3000
3000
2500
2500
2000
2000
1500
1500
1000
1000
500
500
0
0 500 1000 1500 2000 2500 3000 3500 0
0 1000 2000 3000 4000 5000 0 1000
Figure 5.3: The same number of wind turbines are placed when ac-
counting for wake effects as when not. Solution (a) is found after 0.2 seconds
whereas solution (b) is the one obtained after 60 minutes at a MIP-gap of
13.5%.
66
CHAPTER 5. TESTS AND RESULTS
considering the wake effects not yielding a better PARK calculation than
the other is the fact that the Production model over-estimates the production
losses when combining multiple wakes relative to the sum of squares of the
deficits-method [21]. In the configuration of the Figure 5.3(b), the total losses
by wake effects is calculated to be 27% larger than the losses calculated by
PARK. In such tight placements the choice of wake combination model may
thus yield significantly different results.
1,07
1,060
1,06
1,05
1,04
1,033
1,03
1,02
With
wakes
1,01
1,000
1,000
Without
wakes
1,00
0,99
0,98
0,97
100
meter
50
meter
Figure 5.4: Total production with and without wake effects consid-
ered. The number of wind turbines in the 100m and 50m layout is 28 and
30, respectively. The values are scaled with the 100m layout without wakes.
Now, if only optimizing over the total production, placing as many wind tur-
bines as possible into the area will generally be the case also when accounting
for wake effects, since these are small relative to the undisturbed production.
If the placement of yet another wind turbine in a small wind farm yields
a total production gain of merely a fraction of a percent, it is still a gain.
If, however, maximizing the profit, the construction cost of the additional
wind turbine has to be accounted for, which would render the investment
non-profitable. The Production model can be utilized in this manner, and
solutions to the problem with shifting prices of electricity is shown in Figure
5.5.
67
CHAPTER 5. TESTS AND RESULTS
3500
3500
3000
3000
2500
2500
2000
2000
1500 1500
1000 1000
500 500
0 0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
3000
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
Figure 5.5: Solutions for shifting prices of electricity. At low prices only
a few locations are profitable, whereas at high prices even tight placements
are profitable.
68
CHAPTER 5. TESTS AND RESULTS
30
140
Number
of
wind
turbines
25 120
100
20
0
0
0,4
0,45
0,5
0,55
0,6
0,65
0,7
0,75
0,8
SEK/kWh
Figure 5.6: The number of wind turbines and MIP-gap for shifting
prices of electricity. The computing time is 30 min.
shows examples of both, but it is clear that the additional node formulation
is to heavy for the problem to be solved in reasonable time - it had to be
formulated using only one cable type and with edges in only eight directions
to be able to yield any result in half an hour of computation time.
69
CHAPTER 5. TESTS AND RESULTS
3500 3500
13 11 23 20 25 23
1 20 172 1 3000
3000 3
11 7 22 17
2 3 24 2227 24 29 28
2 1 2500
2500 19 12 1
7 4 12 8 1 28 26
3 17 22 2
8 4 26 16
2000 3 7 16 10 3 2000
4 2 4
10 5 18 15
7 5 21 18
2 15 2 2 1 1500
1500 3 62
28 15 3
7 3
6 3 1000
1000 3
9 6
2 500
500 14 9
1
0 0
0 1000 2000 3000 4000 5000 0 1000 0 1000 2000 3000 4000 5000 0 1000
(a) Only edges between wind turbines (b) Additional edges allowed
70
CHAPTER 5. TESTS AND RESULTS
71
CHAPTER 5. TESTS AND RESULTS
4 5 4
4 5 4 4 3 2
4 3 2
(a) The solution found by (b) The solution found by the Produc-
the software Optimize tion model (defined in Section 4.2.6)
solution found, as illustrated in figures 5.8 and 5.9. This is partly similar
to, for example, placing billiard balls at random positions into the triangular
rack. Soon you will not be able to place yet another ball without touching
the ones already placed, and the rack is not filled.
There is also an option in the software Optimize called auto fill, which by no
means performs any form of optimization but is a tool to just fit as many
wind turbines as possible into the project site area. Neither this tool could,
however, place the maximum number of wind turbines in any of the example
problems, as seen in the 50m problem in Figure 5.10.
The comparison
The optimization is carried out using the same possible wind turbine coor-
dinates, project site area, wind turbine type, minimum separation distance,
and wind resource data as in the previous section. Due to the mentioned
drawbacks of the heuristic algorithm, Optimize is not able to fit as many
wind turbines into the project site area as the Production model. Here it
is made sure the number of wind turbines placed before locking their posi-
tions is set higher than the maximum number that can be fitted in the area.
For the 100m grid a total of 18 wind turbines are placed by Optimize and
72
CHAPTER 5. TESTS AND RESULTS
5 4 4 5 4 4
4 3 2 4 3 2
(a) The solution found by the soft- (b) The solution found by the Pro-
ware Optimize duction model (defined in Section
4.2.6)
1,04
1,035
1,032
1,03
0,98
100
meter
50
meter
73
CHAPTER 5. TESTS AND RESULTS
for the 50m grid a total of 20, to be compared to 28 and 30, respectively,
by the Production model. The respective layouts can be viewed in Figure
5.11. Now, performing a PARK calculation for these layouts yields values
of the total production to be compared to the results from the Production
model solutions. The results indicate a 40% higher production for the layouts
found by the Production model, shown in Figure 5.12. A tighter layout will
of course yield a lower efficiency, but the aim of both optimization methods
is to maximize the total production and not the efficiency.
Comparing the total production results of the Optimize locations with the
ones obtained from the Production model when restricted to a maximum of 18
and 20 wind turbines, for the 50m and 100m grid respectively, the Production
model proved to yield an equal or higher production. The total production of
the 50m grid problem, where lack of memory restricted the maximum wake
distance to 900m, was equal to the heuristic Optimize locations. In the 100m
grid problem, the Production model yielded a solution about one percent
better than that of the heuristic Optimize. Total production comparison
for various maximum number of wind turbines is shown in Figure 5.13. The
method of combining multiple wakes in a different way than the “true” PARK
calculation model did thus not yield a worse solution.
74
CHAPTER 5. TESTS AND RESULTS
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 500 1000 1500 2000 2500 3000 3500 0 1000 2000 3000 4000 0 1000 2000
3500
3500
3000
3000
2500
2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 1000 2000 3000 4000 5000 0 1000 0 1000 2000 3000 4000 0 1000 2000
75
CHAPTER 5. TESTS AND RESULTS
1,60
1,43
1,38
1,40
1,20
1,00
1,00
1,00
0,80
Production
model
0,60
WindPRO
Optimize
0,40
0,20
0,00
100
meter
50
meter
1,50
1,40
1,30
Total
production
1,20
0,90
0,80
15
18
21
24
27
Maximum
number
of
wind
turbines
76
CHAPTER 5. TESTS AND RESULTS
5.2 Performance
In this section the speed of the models, using the software and hardware setup
described in the beginning of Chapter 5, are tested. The values of solution
time or MIP-gap would probably change with many of the suggested further
implementations, and should only bee seen as guidance.
The use of wake effect variables highly increases the problem size and perfor-
mance for varying project area and number of nodes is shown in Figure 5.14;
the number of wake variables per number of nodes can be seen in Figure 5.15.
30
25
20
15
10
5
0
0
1
2
3
4
5
6
7
Area
[(km)2]
77
CHAPTER 5. TESTS AND RESULTS
500
integer problem found so far and the currently lowest upper bound of the
continuous problem, is highly increased. This is due to the wake effects,
not being present in the LP relaxation, being of greater importance for the
number of wind turbines placed. Performance of the profit model can be
viewed in Figure 5.16 as well as in Figure 5.6 for the verification example.
78
CHAPTER 5. TESTS AND RESULTS
160 25
100
15
80
60
10
40
5
20
0
0
0,3
0,4
0,5
0,6
0,7
0,8
Electricity
price
[SEK/kWh]
79
CHAPTER 5. TESTS AND RESULTS
Number
of
nodes
[103]
0
2
4
6
8
10
12
14
16
1000,00
100,00
Time
[s]
10,00
1,00
0
5
10
15
20
25
30
35
40
Area
[(km)2]
12
10
MIP
gap
[%]
8
6
4
2
0
0
5
10
15
20
25
30
35
40
Area
[(km)2]
80
CHAPTER 5. TESTS AND RESULTS
The more number of wind turbines to place, the harder the problem to solve.
Figure 5.19 displays the solution time increasing relative to the maximum
number of wind turbines allowed in a 25 (km)2 area.
10000,00
1000,00
Time
[s]
100,00
10,00
1,00
0
20
40
60
80
100
120
Maximum
number
of
wind
turbines
In the Infrastructure model cable and road paths can either be created be-
tween nodes containing wind turbines, or make use of surrounding nodes.
Figures 5.20 and 5.21 present examples of both. Here, computer memory
limitations highly restricted the possible problem sizes.
The simulations performed with the combined model were made — due to
memory limitations — only for really small problem sizes, not interesting
81
CHAPTER 5. TESTS AND RESULTS
29
27
25
MIP
gap
[%]
23
21
122
nodes
257
nodes
19
17
15
4
5
6
7
8
9
10
11
Number
of
wind
turbines
Figure 5.20: Infrastructure model. MIP-gap when using the full grid
of nodes for two problem sizes. The number of paths per node is reduced
to one per sixteen directions. The simulations use variables for two different
cable capacities.
2000
1800
1600
1400
Time
[s]
1200
1000
800
600
400
200
0
0
10
20
30
40
Number
of
nodes
82
CHAPTER 5. TESTS AND RESULTS
enough for this report. It was, however, apparent that due to the small costs
of roads and cables compared to the profit from production, improvements in
the tree costs had a little effect on the MIP-gap in the optimization process.
83
CHAPTER 5. TESTS AND RESULTS
84
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
Chapter 6
85
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
The conclusions listed in this chapter will mainly regard the simulations and
verifications made with the Production and Infrastructure models, whereas
more of the applications and advantages of the models will be referred to in
the discussion.
6.1 Conclusion
The mixed integer linear models derived in this thesis, the Wind Farm Lay-
out Production Optimization model and the Wind Farm Layout Infrastruc-
ture Optimization model, can handle not only production optimization and
a correct combination of noise propagation from multiple wind turbines, but
also the minimization of costs of roads, interconnecting cables and placement
of the transformer station.
The Production model places wind turbines, with respect to maximizing pro-
duction, in an area. It does not miss feasible and good layouts as compared to
the heuristic algorithm used by the commercial software (see Section 5.1.2).
A high resolution is of more importance for finding the optimal production
locations than for minimizing the wake effects. The number of continuous
wake effect variables grows by the square of the number of possible locations;
it can, however, be approximated to grow proportionally to this number by
a suggested approximation of wakes between sectors. Due to the general for-
mulation of sound pressure constraints, the model can incorporate advanced
sound propagation models dependent on wind directions and wind intensi-
ties. The linearity of the model may impose a restriction when combining
wake effects from several wind turbines, since other means of combination
than linear superposition often are proposed. This issue can to some extent
can be compensated by approximations suggested, but this will increase the
problem size. Though not modeling combination of wake effects the same
way as the commercial optimizer, WindPRO 2.6 Optimize, and the total
production calculation software, WindPRO 2.6 PARK, used for finding the
total production, the Production model yielded equal or better solutions than
the commercial software in the verification examples when placing the same
number of wind turbines. When optimizing the total production of the ex-
amples in this thesis, the objective of both the Production model and the
software Optimize, without any limitations on the number of wind turbines,
the Production model was capable of locating positions for about 50% more
installed production capacity than the comparison software. This resulted in
86
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
The combined problem of the Production model and the Infrastructure model
will, however, be large, especially since approximations in order to form a
smaller tree structure cannot be applied here. For this combination to be of
value, the costs of cables and roads should not be just a small fraction of the
revenue of the electricity production. This is because changing the road and
cable costs otherwise will not have much effect on the position of the wind
turbines. A two-stage approach is therefore more likely to be applicable.
The mixed integer linear programs developed in this thesis are capable of
addressing the defined problems of wind farm layout design. The programs
can handle problems of realistic size, but would gain from further improve-
ments. Some of these possible improvements are suggested in the thesis. The
methods used seem to be applicable for use in the industry and compare well
to the commercial software in the verification examples.
87
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
6.2 Discussion
Now, is it possible to gain from the use of computer optimization when de-
signing a wind farm layout? Moreover, is linear programming a prominent
tool for getting more bang for the buck? The answer to both of these ques-
tions would, according to the simulations and discussions in this thesis, be
yes. There are of course many aspects to consider when designing a wind
farm layout, and the human interaction in the workflow is not about to be
replaced by simply pressing an optimization button. However, there seems
to be room for rather large technology improvements regarding the use of
automated optimization.
The comparison between the Production model and the commercial heuristic
software Optimize was made for only a verification example with shifting
grid resolutions, thus the outcome does not have to be general. However, it
illustrates a serious drawback of an heuristic algorithm that places one wind
turbine at a time, and which cannot adjust all of them simultaneously — it
is easy to block combinations of locations resulting in non-optimal solutions.
The integer linear programming approach used in this thesis not only gives
a quality measure during the solution process — the MIP-gap — but is
also generally guaranteed to converge to an optimal solution of the problem
defined. Though it may take a tremendous time and may also require a very
large amount of computer memory. Typically, the optimization process will
be terminated after a certain time or when reaching a certain MIP-gap.
88
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
The Infrastructure model formulation was made under the assumption of the
combination of flows being rather cheap. The implemented model therefore
cannot benefit from placing cables with flows in opposite directions along
the same road, and a closer resemblance to reality is obtained when instead
using the additional formulations suggested. Using additional nodes outside
the fixed wind turbine locations is necessary, forming a so called Steiner tree,
but a full grid of nodes is too heavy to use and the added nodes should be
placed in more motivated locations (as introduced in Section 4.3.7).
89
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
more separated wind turbines does, however, reduce the error of combining
the wake effects linearly.
Now if the project at hand is one where wind turbines are only allowed to
be placed at certain given locations, for example, at shallow sections and
reefs in an offshore project, the Production model avoids the risk of blocking
locations as mentioned above. The exact locations are also easily entered
into the model. Such a problem is, due to the limited number of possible
locations, generally small and additional variables approximating non-linear
combination of wake effects can be added without solution times increasing
to unreasonable levels.
Other types of offshore scenarios, where placement is done more freely but
foundation costs vary greatly with the depth and the type of sea bed, can
also be run in the Production model by optimizing over revenue.
The software Optimize cannot, according to its manual, perform its normal
optimization algorithm on an offshore project or on any flat area at all.
This is because it needs local variations in wind characteristics not to block
too many solutions. Flat areas do not pose a problem for the Production
model. To be able to handle flat areas, Optimize has a mode where it places
wind turbines in a regular grid and then tries varying parameters such as row
spacing to find better solutions. Due to these defined straight or curved rows,
this option is useful for creating visually appealing layouts. This approach
would be possible and fast to implement also in the Production model by
iterating the optimization with changing row parameter values. Due to the
highly restricted number of possible locations in each iteration, the problem
can be solved quickly also with extra variables approximating non-linear
combination of wakes added.
90
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
91
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
depends on whether the locations of the wind turbines are fixed, and both
of these complexity reducing methods ought to be tested.
This thesis has not focused on solution methods, and apart from the prob-
lem specific approximations utilized and suggested, more general means of
relaxations and time reducing techniques should be considered. It may be
possible that Lagrangian relaxation of the constraints relating the wind tur-
bine positions to the tree structure would ease the solving of the combined
problem of the Production model and the Infrastructure model. These con-
straints are however quite complex and other relaxations may prove to be
of greater importance. The choice of branching strategy in the branch and
bound method could also be considered, perhaps as referred to in [5].
92
BIBLIOGRAPHY
Bibliography
[7] C.N. Elkinton, J.F. Manwell, and J.G. McGowan. Offshore wind farm
layout optimization (OWFLO) project: An introduction. In Proceedings
of Copenhagen offshore wind conference 2005, Copenhagen, Denmark,
2005.
93
BIBLIOGRAPHY
[12] S.A. Grady, M.Y. Hussaini, and M.M. Abdullah. Placement of wind
turbines using genetic algorithms. Renewable Energy, 30(2):259–270,
2005.
[14] I. Katić, J. Højstrup, and N.O. Jensen. A simple model for cluster
efficiency. In European Wind Energy Association Conference and Exhi-
bition, pages 407–10, Rome, Italy, 1986.
[15] A. Kusiak and Z. Song. Design of wind farm layout for maximum wind
energy capture. Renewable Energy, 35(3):685–694, 2010.
[20] Hanif D. Sherali and J. Cole Smith. A class o web-based facets for
the generalized vertex packing problem. Discrete Applied Mathematics,
146(3):273–286, 2005.
94
BIBLIOGRAPHY
[22] L.J. Vermeer, J.N. Sørensen, and A. Crespo. Wind turbine wake aero-
dynamics. Progress in Aerospace Sciences, 39(6–7):467–510, 2003.
95